Assessment of multivariate classification analysis methods on 3‐dimensional anatomical distribution of functionally identified cat pre‐Bötzinger complex microinjection sites (712.15)

Abstract

Previous work from our laboratory has demonstrated that multiple patterns of phrenic motor output can be elicited by chemical stimulation of anatomically distinct sub‐regions of the in vivo adult cat pre‐Bötzinger complex. To gain further insight into the anatomical distribution of the microinjection sites yielding these different discharge patterns, we implemented two multivariate classification analysis methods: Fisher Linear Discriminant (FLD) and probabilistic neural network (PNN). For FLD analysis, the multivariate data were transformed into univariate observations by computing the inverse of a pooled variance‐covariance matrix, which was then translated back to the histological coordinates to predict physical sub‐region boundaries. For PNN analysis, a 3‐layered radial basis neural network was created with (1) the input vector for each discharge pattern yielding an output value according to how close the input vector was to each pattern weight vector, (2) the contributions for each class of inputs was summed to produce a net output vector of probabilities; and (3) a transfer function was applied to the outputs to identify the maximum probabilities and produce a classification decision. While each analysis method has its advantages (and disadvantages), both were able to predicte all of the observed phrenic discharge patterns based solely on the 3‐D coordinates of the microinjection sites. We suggest, however, that FLD analysis may be better suited for these data.Grant Funding Source: Supported by HL63175